I’ve been watching this pattern emerge for years now, and it’s finally crystallizing into something I can’t ignore.
Back in May 2018, and that feels like yesterday, honestly when Dialpad acquired TalkIQ, I wrote that VoiceAI would become as ubiquitous as collaboration itself. What struck me then wasn’t just that Dialpad was adding artificial intelligence to phone calls. It was that they were building intelligence around conversations, something a general-purpose AI couldn’t accomplish without truly understanding the nuances of human communication.
Eight years later, that insight is becoming the blueprint for AI’s next evolution — and it’s not what most people expect.
Here’s the thing: the future isn’t another giant LLM trying to know everything about everything. It’s specialized AI that knows something exceptionally well. And I mean exceptionally.
Take Snapp, which just raised $4.6 million with backing from a16z Speedrun. It’s a sports-focused AI that combines live data, predictive analytics, and domain intelligence. They’re claiming 44 percent greater accuracy than general-purpose models on their internal benchmarks. Now, that remains independently unverified — and I’m always skeptical of self-reported metrics — but the strategic premise is absolutely compelling.
Think about it this way: sports doesn’t need an AI that can write poetry or explain quantum mechanics. It needs one that knows who’s injured, who’s playing, what the weather conditions are, and what’s happening right now. That’s a fundamentally different value proposition.
We’ve seen this movie before, haven’t we?
In 2025, when I was writing about vCONs, I argued that conversational context was becoming the new currency in telecommunications. In my analysis of telephony-first AI, I warned that startups treating voice merely as data, just another input stream, were missing the entire communications foundation that makes voice unique. And in June, I wrote that the model isn’t the product; the output is.
The pattern is unmistakable once you start looking for it.
Generic LLMs provide intelligence: broad, shallow, impressive in its breadth. Specialized systems provide expertise, context, timely data, and actionable outcomes. That’s not a subtle distinction. That’s the difference between a reference library and a trusted advisor who knows your business inside and out.
And that’s the distinction investors should be examining with a magnifying glass.
The winners in this next phase won’t necessarily train their own foundation models from scratch; that’s an expensive, time-consuming endeavor that requires resources most startups simply don’t have. Instead, they’ll combine powerful existing models with proprietary information, specialized reasoning capabilities, workflow integration, and domain-specific experience that can’t be replicated by throwing more compute at the problem.
Dialpad understood this in 2018, when the rest of the world was still figuring out what AI even meant in practical terms. Snapp is betting on it in 2026. And I suspect we’re going to see this pattern repeat across dozens of verticals, including healthcare, legal, finance, manufacturing, you name it.
Here’s what I keep coming back to: the next AI winners won’t be the companies that know the most about everything. They’ll be the companies that know exactly what matters, when it matters, and what to do about it.
That last part, “what to do about it” is crucial. It’s not enough to have information or even insights. You need actionable intelligence that fits into existing workflows and solves real problems that people are willing to pay to solve.
This reminds me of the early days of blogging, actually. The platforms that won weren’t the ones with the most features or the most sophisticated technology. They were the ones that understood what bloggers actually needed to do: publish quickly, share easily, and connect with readers. The technology served the use case, not the other way around.
We’re at a similar inflection point with AI. The foundation models, your GPT-4s and Claudes and Geminis, are like the early internet infrastructure. Necessary, powerful, but ultimately commoditized. The real value creation happens in the layer above: the specialized applications that take that raw capability and channel it into something specific, something useful, something that solves a problem you didn’t even realize could be solved.
And that’s where the real value gets created. Not in the models themselves, but in the expertise, data pipelines, domain knowledge, and workflow integration that turn a general-purpose tool into an indispensable specialist.
I’m watching this space closely, and I think we’re just getting started. The AI gold rush everyone’s been talking about? It’s not about scale anymore. It’s about precision, context, and knowing your domain better than anyone else.
That’s the game now.